Using Artificial Neural Network and Machine Learning to predict magnetic alloys behavior.

Abstract

Magnetic refrigeration, based on the magnetocaloric effect, offers a promising alternative to conventional refrigeration due to its high efficiency and environmental benefits. Discovering magnetocaloric materials is important for this technology but it time consuming and costly to do it the traditional way. This study uses Artificial Neural Networks (ANN) and machine learning (ML) models to predict properties of magnetic alloys, i.e., Curie temperature (Tc), magnetic field change (H) and magnetic entropy change (ΔS), from their compositions.  

   A dataset of 200 compositions, including atomic percentage values of elements like Gd, Dy, and Tb, etc, along with H, and Tc as input and ΔS as the output. Data preprocessing included normalization and handling missing values. The algorithms used were Random Forest (RF), Gradient Boosting Regressor (GBR), Linear Regression (LR), and an ANN consisting of three hidden layers with ReLU activation and Adam optimizer. Metrics: Loss, MSE, R², and MAE. 

   The result show that GBR outperformed other model with test MSE (20.87), MAE (3.23), and R² (0.865). RFR performed moderately with training MSE (9.37), R² (0.94), test MSE (20.31), R² (0.87)., while ANN showed strong training performance but struggled with overall training MSE (0.49), MAE (0.43), and loss (0.49). LR performed the least with a test MSE of 61.80 and MAE of 5.18.  

   These findings show the potential of machine learning for the prediction of magnetocaloric properties and further refinement and additional data will improve model performance and aid in the discovery of new magnetic materials.